1 //===- GVN.cpp - Eliminate redundant values and loads ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
13 // Note that this pass does the value numbering itself, it does not use the
14 // ValueNumbering analysis passes.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "gvn"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/BasicBlock.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Value.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/DepthFirstIterator.h"
28 #include "llvm/ADT/SmallPtrSet.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/Dominators.h"
32 #include "llvm/Analysis/AliasAnalysis.h"
33 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
34 #include "llvm/Support/CFG.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Compiler.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
42 STATISTIC(NumGVNInstr, "Number of instructions deleted");
43 STATISTIC(NumGVNLoad, "Number of loads deleted");
44 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
45 STATISTIC(NumGVNBlocks, "Number of blocks merged");
47 static cl::opt<bool> EnablePRE("enable-pre",
48 cl::init(true), cl::Hidden);
50 //===----------------------------------------------------------------------===//
52 //===----------------------------------------------------------------------===//
54 /// This class holds the mapping between values and value numbers. It is used
55 /// as an efficient mechanism to determine the expression-wise equivalence of
58 struct VISIBILITY_HIDDEN Expression {
59 enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM,
60 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
61 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
62 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
63 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
64 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
65 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
66 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
67 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
68 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
71 ExpressionOpcode opcode;
76 SmallVector<uint32_t, 4> varargs;
80 Expression(ExpressionOpcode o) : opcode(o) { }
82 bool operator==(const Expression &other) const {
83 if (opcode != other.opcode)
85 else if (opcode == EMPTY || opcode == TOMBSTONE)
87 else if (type != other.type)
89 else if (function != other.function)
91 else if (firstVN != other.firstVN)
93 else if (secondVN != other.secondVN)
95 else if (thirdVN != other.thirdVN)
98 if (varargs.size() != other.varargs.size())
101 for (size_t i = 0; i < varargs.size(); ++i)
102 if (varargs[i] != other.varargs[i])
109 bool operator!=(const Expression &other) const {
110 if (opcode != other.opcode)
112 else if (opcode == EMPTY || opcode == TOMBSTONE)
114 else if (type != other.type)
116 else if (function != other.function)
118 else if (firstVN != other.firstVN)
120 else if (secondVN != other.secondVN)
122 else if (thirdVN != other.thirdVN)
125 if (varargs.size() != other.varargs.size())
128 for (size_t i = 0; i < varargs.size(); ++i)
129 if (varargs[i] != other.varargs[i])
137 class VISIBILITY_HIDDEN ValueTable {
139 DenseMap<Value*, uint32_t> valueNumbering;
140 DenseMap<Expression, uint32_t> expressionNumbering;
142 MemoryDependenceAnalysis* MD;
145 uint32_t nextValueNumber;
147 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
148 Expression::ExpressionOpcode getOpcode(CmpInst* C);
149 Expression::ExpressionOpcode getOpcode(CastInst* C);
150 Expression create_expression(BinaryOperator* BO);
151 Expression create_expression(CmpInst* C);
152 Expression create_expression(ShuffleVectorInst* V);
153 Expression create_expression(ExtractElementInst* C);
154 Expression create_expression(InsertElementInst* V);
155 Expression create_expression(SelectInst* V);
156 Expression create_expression(CastInst* C);
157 Expression create_expression(GetElementPtrInst* G);
158 Expression create_expression(CallInst* C);
159 Expression create_expression(Constant* C);
161 ValueTable() : nextValueNumber(1) { }
162 uint32_t lookup_or_add(Value* V);
163 uint32_t lookup(Value* V) const;
164 void add(Value* V, uint32_t num);
166 void erase(Value* v);
168 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
169 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
170 void setDomTree(DominatorTree* D) { DT = D; }
171 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
176 template <> struct DenseMapInfo<Expression> {
177 static inline Expression getEmptyKey() {
178 return Expression(Expression::EMPTY);
181 static inline Expression getTombstoneKey() {
182 return Expression(Expression::TOMBSTONE);
185 static unsigned getHashValue(const Expression e) {
186 unsigned hash = e.opcode;
188 hash = e.firstVN + hash * 37;
189 hash = e.secondVN + hash * 37;
190 hash = e.thirdVN + hash * 37;
192 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
193 (unsigned)((uintptr_t)e.type >> 9)) +
196 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
197 E = e.varargs.end(); I != E; ++I)
198 hash = *I + hash * 37;
200 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
201 (unsigned)((uintptr_t)e.function >> 9)) +
206 static bool isEqual(const Expression &LHS, const Expression &RHS) {
209 static bool isPod() { return true; }
213 //===----------------------------------------------------------------------===//
214 // ValueTable Internal Functions
215 //===----------------------------------------------------------------------===//
216 Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
217 switch(BO->getOpcode()) {
218 default: // THIS SHOULD NEVER HAPPEN
219 assert(0 && "Binary operator with unknown opcode?");
220 case Instruction::Add: return Expression::ADD;
221 case Instruction::Sub: return Expression::SUB;
222 case Instruction::Mul: return Expression::MUL;
223 case Instruction::UDiv: return Expression::UDIV;
224 case Instruction::SDiv: return Expression::SDIV;
225 case Instruction::FDiv: return Expression::FDIV;
226 case Instruction::URem: return Expression::UREM;
227 case Instruction::SRem: return Expression::SREM;
228 case Instruction::FRem: return Expression::FREM;
229 case Instruction::Shl: return Expression::SHL;
230 case Instruction::LShr: return Expression::LSHR;
231 case Instruction::AShr: return Expression::ASHR;
232 case Instruction::And: return Expression::AND;
233 case Instruction::Or: return Expression::OR;
234 case Instruction::Xor: return Expression::XOR;
238 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
239 if (isa<ICmpInst>(C) || isa<VICmpInst>(C)) {
240 switch (C->getPredicate()) {
241 default: // THIS SHOULD NEVER HAPPEN
242 assert(0 && "Comparison with unknown predicate?");
243 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
244 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
245 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
246 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
247 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
248 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
249 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
250 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
251 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
252 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
255 assert((isa<FCmpInst>(C) || isa<VFCmpInst>(C)) && "Unknown compare");
256 switch (C->getPredicate()) {
257 default: // THIS SHOULD NEVER HAPPEN
258 assert(0 && "Comparison with unknown predicate?");
259 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
260 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
261 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
262 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
263 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
264 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
265 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
266 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
267 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
268 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
269 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
270 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
271 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
272 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
276 Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
277 switch(C->getOpcode()) {
278 default: // THIS SHOULD NEVER HAPPEN
279 assert(0 && "Cast operator with unknown opcode?");
280 case Instruction::Trunc: return Expression::TRUNC;
281 case Instruction::ZExt: return Expression::ZEXT;
282 case Instruction::SExt: return Expression::SEXT;
283 case Instruction::FPToUI: return Expression::FPTOUI;
284 case Instruction::FPToSI: return Expression::FPTOSI;
285 case Instruction::UIToFP: return Expression::UITOFP;
286 case Instruction::SIToFP: return Expression::SITOFP;
287 case Instruction::FPTrunc: return Expression::FPTRUNC;
288 case Instruction::FPExt: return Expression::FPEXT;
289 case Instruction::PtrToInt: return Expression::PTRTOINT;
290 case Instruction::IntToPtr: return Expression::INTTOPTR;
291 case Instruction::BitCast: return Expression::BITCAST;
295 Expression ValueTable::create_expression(CallInst* C) {
298 e.type = C->getType();
302 e.function = C->getCalledFunction();
303 e.opcode = Expression::CALL;
305 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
307 e.varargs.push_back(lookup_or_add(*I));
312 Expression ValueTable::create_expression(BinaryOperator* BO) {
315 e.firstVN = lookup_or_add(BO->getOperand(0));
316 e.secondVN = lookup_or_add(BO->getOperand(1));
319 e.type = BO->getType();
320 e.opcode = getOpcode(BO);
325 Expression ValueTable::create_expression(CmpInst* C) {
328 e.firstVN = lookup_or_add(C->getOperand(0));
329 e.secondVN = lookup_or_add(C->getOperand(1));
332 e.type = C->getType();
333 e.opcode = getOpcode(C);
338 Expression ValueTable::create_expression(CastInst* C) {
341 e.firstVN = lookup_or_add(C->getOperand(0));
345 e.type = C->getType();
346 e.opcode = getOpcode(C);
351 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
354 e.firstVN = lookup_or_add(S->getOperand(0));
355 e.secondVN = lookup_or_add(S->getOperand(1));
356 e.thirdVN = lookup_or_add(S->getOperand(2));
358 e.type = S->getType();
359 e.opcode = Expression::SHUFFLE;
364 Expression ValueTable::create_expression(ExtractElementInst* E) {
367 e.firstVN = lookup_or_add(E->getOperand(0));
368 e.secondVN = lookup_or_add(E->getOperand(1));
371 e.type = E->getType();
372 e.opcode = Expression::EXTRACT;
377 Expression ValueTable::create_expression(InsertElementInst* I) {
380 e.firstVN = lookup_or_add(I->getOperand(0));
381 e.secondVN = lookup_or_add(I->getOperand(1));
382 e.thirdVN = lookup_or_add(I->getOperand(2));
384 e.type = I->getType();
385 e.opcode = Expression::INSERT;
390 Expression ValueTable::create_expression(SelectInst* I) {
393 e.firstVN = lookup_or_add(I->getCondition());
394 e.secondVN = lookup_or_add(I->getTrueValue());
395 e.thirdVN = lookup_or_add(I->getFalseValue());
397 e.type = I->getType();
398 e.opcode = Expression::SELECT;
403 Expression ValueTable::create_expression(GetElementPtrInst* G) {
406 e.firstVN = lookup_or_add(G->getPointerOperand());
410 e.type = G->getType();
411 e.opcode = Expression::GEP;
413 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
415 e.varargs.push_back(lookup_or_add(*I));
420 //===----------------------------------------------------------------------===//
421 // ValueTable External Functions
422 //===----------------------------------------------------------------------===//
424 /// add - Insert a value into the table with a specified value number.
425 void ValueTable::add(Value* V, uint32_t num) {
426 valueNumbering.insert(std::make_pair(V, num));
429 /// lookup_or_add - Returns the value number for the specified value, assigning
430 /// it a new number if it did not have one before.
431 uint32_t ValueTable::lookup_or_add(Value* V) {
432 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
433 if (VI != valueNumbering.end())
436 if (CallInst* C = dyn_cast<CallInst>(V)) {
437 if (AA->doesNotAccessMemory(C)) {
438 Expression e = create_expression(C);
440 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
441 if (EI != expressionNumbering.end()) {
442 valueNumbering.insert(std::make_pair(V, EI->second));
445 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
446 valueNumbering.insert(std::make_pair(V, nextValueNumber));
448 return nextValueNumber++;
450 } else if (AA->onlyReadsMemory(C)) {
451 Expression e = create_expression(C);
453 if (expressionNumbering.find(e) == expressionNumbering.end()) {
454 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
455 valueNumbering.insert(std::make_pair(V, nextValueNumber));
456 return nextValueNumber++;
459 Instruction* local_dep = MD->getDependency(C);
461 if (local_dep == MemoryDependenceAnalysis::None) {
462 valueNumbering.insert(std::make_pair(V, nextValueNumber));
463 return nextValueNumber++;
464 } else if (local_dep != MemoryDependenceAnalysis::NonLocal) {
465 if (!isa<CallInst>(local_dep)) {
466 valueNumbering.insert(std::make_pair(V, nextValueNumber));
467 return nextValueNumber++;
470 CallInst* local_cdep = cast<CallInst>(local_dep);
472 if (local_cdep->getCalledFunction() != C->getCalledFunction() ||
473 local_cdep->getNumOperands() != C->getNumOperands()) {
474 valueNumbering.insert(std::make_pair(V, nextValueNumber));
475 return nextValueNumber++;
476 } else if (!C->getCalledFunction()) {
477 valueNumbering.insert(std::make_pair(V, nextValueNumber));
478 return nextValueNumber++;
480 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
481 uint32_t c_vn = lookup_or_add(C->getOperand(i));
482 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
484 valueNumbering.insert(std::make_pair(V, nextValueNumber));
485 return nextValueNumber++;
489 uint32_t v = lookup_or_add(local_cdep);
490 valueNumbering.insert(std::make_pair(V, v));
496 DenseMap<BasicBlock*, Value*> deps;
497 MD->getNonLocalDependency(C, deps);
500 for (DenseMap<BasicBlock*, Value*>::iterator I = deps.begin(),
501 E = deps.end(); I != E; ++I) {
502 if (I->second == MemoryDependenceAnalysis::None) {
503 valueNumbering.insert(std::make_pair(V, nextValueNumber));
505 return nextValueNumber++;
506 } else if (I->second != MemoryDependenceAnalysis::NonLocal) {
507 if (DT->properlyDominates(I->first, C->getParent())) {
508 if (CallInst* CD = dyn_cast<CallInst>(I->second))
511 valueNumbering.insert(std::make_pair(V, nextValueNumber));
512 return nextValueNumber++;
515 valueNumbering.insert(std::make_pair(V, nextValueNumber));
516 return nextValueNumber++;
522 valueNumbering.insert(std::make_pair(V, nextValueNumber));
523 return nextValueNumber++;
526 if (cdep->getCalledFunction() != C->getCalledFunction() ||
527 cdep->getNumOperands() != C->getNumOperands()) {
528 valueNumbering.insert(std::make_pair(V, nextValueNumber));
529 return nextValueNumber++;
530 } else if (!C->getCalledFunction()) {
531 valueNumbering.insert(std::make_pair(V, nextValueNumber));
532 return nextValueNumber++;
534 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
535 uint32_t c_vn = lookup_or_add(C->getOperand(i));
536 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
538 valueNumbering.insert(std::make_pair(V, nextValueNumber));
539 return nextValueNumber++;
543 uint32_t v = lookup_or_add(cdep);
544 valueNumbering.insert(std::make_pair(V, v));
549 valueNumbering.insert(std::make_pair(V, nextValueNumber));
550 return nextValueNumber++;
552 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
553 Expression e = create_expression(BO);
555 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
556 if (EI != expressionNumbering.end()) {
557 valueNumbering.insert(std::make_pair(V, EI->second));
560 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
561 valueNumbering.insert(std::make_pair(V, nextValueNumber));
563 return nextValueNumber++;
565 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) {
566 Expression e = create_expression(C);
568 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
569 if (EI != expressionNumbering.end()) {
570 valueNumbering.insert(std::make_pair(V, EI->second));
573 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
574 valueNumbering.insert(std::make_pair(V, nextValueNumber));
576 return nextValueNumber++;
578 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
579 Expression e = create_expression(U);
581 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
582 if (EI != expressionNumbering.end()) {
583 valueNumbering.insert(std::make_pair(V, EI->second));
586 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
587 valueNumbering.insert(std::make_pair(V, nextValueNumber));
589 return nextValueNumber++;
591 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) {
592 Expression e = create_expression(U);
594 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
595 if (EI != expressionNumbering.end()) {
596 valueNumbering.insert(std::make_pair(V, EI->second));
599 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
600 valueNumbering.insert(std::make_pair(V, nextValueNumber));
602 return nextValueNumber++;
604 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) {
605 Expression e = create_expression(U);
607 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
608 if (EI != expressionNumbering.end()) {
609 valueNumbering.insert(std::make_pair(V, EI->second));
612 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
613 valueNumbering.insert(std::make_pair(V, nextValueNumber));
615 return nextValueNumber++;
617 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) {
618 Expression e = create_expression(U);
620 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
621 if (EI != expressionNumbering.end()) {
622 valueNumbering.insert(std::make_pair(V, EI->second));
625 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
626 valueNumbering.insert(std::make_pair(V, nextValueNumber));
628 return nextValueNumber++;
630 } else if (CastInst* U = dyn_cast<CastInst>(V)) {
631 Expression e = create_expression(U);
633 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
634 if (EI != expressionNumbering.end()) {
635 valueNumbering.insert(std::make_pair(V, EI->second));
638 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
639 valueNumbering.insert(std::make_pair(V, nextValueNumber));
641 return nextValueNumber++;
643 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
644 Expression e = create_expression(U);
646 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
647 if (EI != expressionNumbering.end()) {
648 valueNumbering.insert(std::make_pair(V, EI->second));
651 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
652 valueNumbering.insert(std::make_pair(V, nextValueNumber));
654 return nextValueNumber++;
657 valueNumbering.insert(std::make_pair(V, nextValueNumber));
658 return nextValueNumber++;
662 /// lookup - Returns the value number of the specified value. Fails if
663 /// the value has not yet been numbered.
664 uint32_t ValueTable::lookup(Value* V) const {
665 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
666 assert(VI != valueNumbering.end() && "Value not numbered?");
670 /// clear - Remove all entries from the ValueTable
671 void ValueTable::clear() {
672 valueNumbering.clear();
673 expressionNumbering.clear();
677 /// erase - Remove a value from the value numbering
678 void ValueTable::erase(Value* V) {
679 valueNumbering.erase(V);
682 //===----------------------------------------------------------------------===//
684 //===----------------------------------------------------------------------===//
687 struct VISIBILITY_HIDDEN ValueNumberScope {
688 ValueNumberScope* parent;
689 DenseMap<uint32_t, Value*> table;
691 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
697 class VISIBILITY_HIDDEN GVN : public FunctionPass {
698 bool runOnFunction(Function &F);
700 static char ID; // Pass identification, replacement for typeid
701 GVN() : FunctionPass(&ID) { }
705 DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
707 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
711 // This transformation requires dominator postdominator info
712 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
713 AU.addRequired<DominatorTree>();
714 AU.addRequired<MemoryDependenceAnalysis>();
715 AU.addRequired<AliasAnalysis>();
717 AU.addPreserved<DominatorTree>();
718 AU.addPreserved<AliasAnalysis>();
722 // FIXME: eliminate or document these better
723 bool processLoad(LoadInst* L,
724 DenseMap<Value*, LoadInst*> &lastLoad,
725 SmallVectorImpl<Instruction*> &toErase);
726 bool processInstruction(Instruction* I,
727 DenseMap<Value*, LoadInst*>& lastSeenLoad,
728 SmallVectorImpl<Instruction*> &toErase);
729 bool processNonLocalLoad(LoadInst* L,
730 SmallVectorImpl<Instruction*> &toErase);
731 bool processBlock(DomTreeNode* DTN);
732 Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig,
733 DenseMap<BasicBlock*, Value*> &Phis,
734 bool top_level = false);
735 void dump(DenseMap<uint32_t, Value*>& d);
736 bool iterateOnFunction(Function &F);
737 Value* CollapsePhi(PHINode* p);
738 bool isSafeReplacement(PHINode* p, Instruction* inst);
739 bool performPRE(Function& F);
740 Value* lookupNumber(BasicBlock* BB, uint32_t num);
741 bool mergeBlockIntoPredecessor(BasicBlock* BB);
742 void cleanupGlobalSets();
748 // createGVNPass - The public interface to this file...
749 FunctionPass *llvm::createGVNPass() { return new GVN(); }
751 static RegisterPass<GVN> X("gvn",
752 "Global Value Numbering");
754 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
756 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
757 E = d.end(); I != E; ++I) {
758 printf("%d\n", I->first);
764 Value* GVN::CollapsePhi(PHINode* p) {
765 DominatorTree &DT = getAnalysis<DominatorTree>();
766 Value* constVal = p->hasConstantValue();
768 if (!constVal) return 0;
770 Instruction* inst = dyn_cast<Instruction>(constVal);
774 if (DT.dominates(inst, p))
775 if (isSafeReplacement(p, inst))
780 bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) {
781 if (!isa<PHINode>(inst))
784 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
786 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
787 if (use_phi->getParent() == inst->getParent())
793 /// GetValueForBlock - Get the value to use within the specified basic block.
794 /// available values are in Phis.
795 Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig,
796 DenseMap<BasicBlock*, Value*> &Phis,
799 // If we have already computed this value, return the previously computed val.
800 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
801 if (V != Phis.end() && !top_level) return V->second;
803 // If the block is unreachable, just return undef, since this path
804 // can't actually occur at runtime.
805 if (!getAnalysis<DominatorTree>().isReachableFromEntry(BB))
806 return Phis[BB] = UndefValue::get(orig->getType());
808 BasicBlock* singlePred = BB->getSinglePredecessor();
810 Value *ret = GetValueForBlock(singlePred, orig, Phis);
815 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so
816 // now, then get values to fill in the incoming values for the PHI.
817 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle",
819 PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB)));
821 if (Phis.count(BB) == 0)
822 Phis.insert(std::make_pair(BB, PN));
824 // Fill in the incoming values for the block.
825 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
826 Value* val = GetValueForBlock(*PI, orig, Phis);
827 PN->addIncoming(val, *PI);
830 AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
831 AA.copyValue(orig, PN);
833 // Attempt to collapse PHI nodes that are trivially redundant
834 Value* v = CollapsePhi(PN);
836 // Cache our phi construction results
837 phiMap[orig->getPointerOperand()].insert(PN);
841 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
843 MD.removeInstruction(PN);
844 PN->replaceAllUsesWith(v);
846 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
847 E = Phis.end(); I != E; ++I)
851 PN->eraseFromParent();
857 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
858 /// non-local by performing PHI construction.
859 bool GVN::processNonLocalLoad(LoadInst* L,
860 SmallVectorImpl<Instruction*> &toErase) {
861 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
863 // Find the non-local dependencies of the load
864 DenseMap<BasicBlock*, Value*> deps;
865 MD.getNonLocalDependency(L, deps);
867 // If we had to process more than one hundred blocks to find the
868 // dependencies, this load isn't worth worrying about. Optimizing
869 // it will be too expensive.
870 if (deps.size() > 100)
873 DenseMap<BasicBlock*, Value*> repl;
875 // Filter out useless results (non-locals, etc)
876 for (DenseMap<BasicBlock*, Value*>::iterator I = deps.begin(), E = deps.end();
878 if (I->second == MemoryDependenceAnalysis::None)
881 if (I->second == MemoryDependenceAnalysis::NonLocal)
884 if (StoreInst* S = dyn_cast<StoreInst>(I->second)) {
885 if (S->getPointerOperand() != L->getPointerOperand())
887 repl[I->first] = S->getOperand(0);
888 } else if (LoadInst* LD = dyn_cast<LoadInst>(I->second)) {
889 if (LD->getPointerOperand() != L->getPointerOperand())
897 // Use cached PHI construction information from previous runs
898 SmallPtrSet<Instruction*, 4>& p = phiMap[L->getPointerOperand()];
899 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
901 if ((*I)->getParent() == L->getParent()) {
902 MD.removeInstruction(L);
903 L->replaceAllUsesWith(*I);
904 toErase.push_back(L);
909 repl.insert(std::make_pair((*I)->getParent(), *I));
912 // Perform PHI construction
913 SmallPtrSet<BasicBlock*, 4> visited;
914 Value* v = GetValueForBlock(L->getParent(), L, repl, true);
916 MD.removeInstruction(L);
917 L->replaceAllUsesWith(v);
918 toErase.push_back(L);
924 /// processLoad - Attempt to eliminate a load, first by eliminating it
925 /// locally, and then attempting non-local elimination if that fails.
926 bool GVN::processLoad(LoadInst *L, DenseMap<Value*, LoadInst*> &lastLoad,
927 SmallVectorImpl<Instruction*> &toErase) {
928 if (L->isVolatile()) {
929 lastLoad[L->getPointerOperand()] = L;
933 Value* pointer = L->getPointerOperand();
934 LoadInst*& last = lastLoad[pointer];
936 // ... to a pointer that has been loaded from before...
937 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
938 bool removedNonLocal = false;
939 Instruction* dep = MD.getDependency(L);
940 if (dep == MemoryDependenceAnalysis::NonLocal &&
941 L->getParent() != &L->getParent()->getParent()->getEntryBlock()) {
942 removedNonLocal = processNonLocalLoad(L, toErase);
944 if (!removedNonLocal)
947 return removedNonLocal;
951 bool deletedLoad = false;
953 // Walk up the dependency chain until we either find
954 // a dependency we can use, or we can't walk any further
955 while (dep != MemoryDependenceAnalysis::None &&
956 dep != MemoryDependenceAnalysis::NonLocal &&
957 (isa<LoadInst>(dep) || isa<StoreInst>(dep))) {
958 // ... that depends on a store ...
959 if (StoreInst* S = dyn_cast<StoreInst>(dep)) {
960 if (S->getPointerOperand() == pointer) {
962 MD.removeInstruction(L);
964 L->replaceAllUsesWith(S->getOperand(0));
965 toErase.push_back(L);
970 // Whether we removed it or not, we can't
974 // If we don't depend on a store, and we haven't
975 // been loaded before, bail.
977 } else if (dep == last) {
979 MD.removeInstruction(L);
981 L->replaceAllUsesWith(last);
982 toErase.push_back(L);
988 dep = MD.getDependency(L, dep);
992 if (dep != MemoryDependenceAnalysis::None &&
993 dep != MemoryDependenceAnalysis::NonLocal &&
994 isa<AllocationInst>(dep)) {
995 // Check that this load is actually from the
996 // allocation we found
997 if (L->getOperand(0)->getUnderlyingObject() == dep) {
998 // If this load depends directly on an allocation, there isn't
999 // anything stored there; therefore, we can optimize this load
1001 MD.removeInstruction(L);
1003 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1004 toErase.push_back(L);
1016 Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) {
1017 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
1018 if (I == localAvail.end())
1021 ValueNumberScope* locals = I->second;
1024 DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num);
1025 if (I != locals->table.end())
1028 locals = locals->parent;
1034 /// processInstruction - When calculating availability, handle an instruction
1035 /// by inserting it into the appropriate sets
1036 bool GVN::processInstruction(Instruction *I,
1037 DenseMap<Value*, LoadInst*> &lastSeenLoad,
1038 SmallVectorImpl<Instruction*> &toErase) {
1039 if (LoadInst* L = dyn_cast<LoadInst>(I)) {
1040 bool changed = processLoad(L, lastSeenLoad, toErase);
1043 unsigned num = VN.lookup_or_add(L);
1044 localAvail[I->getParent()]->table.insert(std::make_pair(num, L));
1050 uint32_t nextNum = VN.getNextUnusedValueNumber();
1051 unsigned num = VN.lookup_or_add(I);
1053 // Allocations are always uniquely numbered, so we can save time and memory
1054 // by fast failing them.
1055 if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) {
1056 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1060 // Collapse PHI nodes
1061 if (PHINode* p = dyn_cast<PHINode>(I)) {
1062 Value* constVal = CollapsePhi(p);
1065 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
1067 if (PI->second.count(p))
1068 PI->second.erase(p);
1070 p->replaceAllUsesWith(constVal);
1071 toErase.push_back(p);
1073 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1076 // If the number we were assigned was a brand new VN, then we don't
1077 // need to do a lookup to see if the number already exists
1078 // somewhere in the domtree: it can't!
1079 } else if (num == nextNum) {
1080 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1082 // Perform value-number based elimination
1083 } else if (Value* repl = lookupNumber(I->getParent(), num)) {
1085 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
1086 MD.removeInstruction(I);
1089 I->replaceAllUsesWith(repl);
1090 toErase.push_back(I);
1093 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1099 // GVN::runOnFunction - This is the main transformation entry point for a
1102 bool GVN::runOnFunction(Function& F) {
1103 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1104 VN.setMemDep(&getAnalysis<MemoryDependenceAnalysis>());
1105 VN.setDomTree(&getAnalysis<DominatorTree>());
1107 bool changed = false;
1108 bool shouldContinue = true;
1110 // Merge unconditional branches, allowing PRE to catch more
1111 // optimization opportunities.
1112 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1113 BasicBlock* BB = FI;
1115 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1116 if (removedBlock) NumGVNBlocks++;
1118 changed |= removedBlock;
1121 while (shouldContinue) {
1122 shouldContinue = iterateOnFunction(F);
1123 changed |= shouldContinue;
1127 bool PREChanged = true;
1128 while (PREChanged) {
1129 PREChanged = performPRE(F);
1130 changed |= PREChanged;
1134 cleanupGlobalSets();
1140 bool GVN::processBlock(DomTreeNode* DTN) {
1141 BasicBlock* BB = DTN->getBlock();
1143 SmallVector<Instruction*, 8> toErase;
1144 DenseMap<Value*, LoadInst*> lastSeenLoad;
1145 bool changed_function = false;
1149 new ValueNumberScope(localAvail[DTN->getIDom()->getBlock()]);
1151 localAvail[BB] = new ValueNumberScope(0);
1153 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1155 changed_function |= processInstruction(BI, lastSeenLoad, toErase);
1156 if (toErase.empty()) {
1161 // If we need some instructions deleted, do it now.
1162 NumGVNInstr += toErase.size();
1164 // Avoid iterator invalidation.
1165 bool AtStart = BI == BB->begin();
1169 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1170 E = toErase.end(); I != E; ++I)
1171 (*I)->eraseFromParent();
1181 return changed_function;
1184 /// performPRE - Perform a purely local form of PRE that looks for diamond
1185 /// control flow patterns and attempts to perform simple PRE at the join point.
1186 bool GVN::performPRE(Function& F) {
1187 bool changed = false;
1188 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
1189 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
1190 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
1191 BasicBlock* CurrentBlock = *DI;
1193 // Nothing to PRE in the entry block.
1194 if (CurrentBlock == &F.getEntryBlock()) continue;
1196 for (BasicBlock::iterator BI = CurrentBlock->begin(),
1197 BE = CurrentBlock->end(); BI != BE; ) {
1198 if (isa<AllocationInst>(BI) || isa<TerminatorInst>(BI) ||
1199 isa<PHINode>(BI) || BI->mayReadFromMemory() ||
1200 BI->mayWriteToMemory()) {
1205 uint32_t valno = VN.lookup(BI);
1207 // Look for the predecessors for PRE opportunities. We're
1208 // only trying to solve the basic diamond case, where
1209 // a value is computed in the successor and one predecessor,
1210 // but not the other. We also explicitly disallow cases
1211 // where the successor is its own predecessor, because they're
1212 // more complicated to get right.
1213 unsigned numWith = 0;
1214 unsigned numWithout = 0;
1215 BasicBlock* PREPred = 0;
1216 DenseMap<BasicBlock*, Value*> predMap;
1217 for (pred_iterator PI = pred_begin(CurrentBlock),
1218 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
1219 // We're not interested in PRE where the block is its
1220 // own predecessor, on in blocks with predecessors
1221 // that are not reachable.
1222 if (*PI == CurrentBlock) {
1225 } else if (!localAvail.count(*PI)) {
1230 DenseMap<uint32_t, Value*>::iterator predV =
1231 localAvail[*PI]->table.find(valno);
1232 if (predV == localAvail[*PI]->table.end()) {
1235 } else if (predV->second == BI) {
1238 predMap[*PI] = predV->second;
1243 // Don't do PRE when it might increase code size, i.e. when
1244 // we would need to insert instructions in more than one pred.
1245 if (numWithout != 1 || numWith == 0) {
1250 // We can't do PRE safely on a critical edge, so instead we schedule
1251 // the edge to be split and perform the PRE the next time we iterate
1253 unsigned succNum = 0;
1254 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
1256 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) {
1261 if (isCriticalEdge(PREPred->getTerminator(), succNum)) {
1262 toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum));
1268 // Instantiate the expression the in predecessor that lacked it.
1269 // Because we are going top-down through the block, all value numbers
1270 // will be available in the predecessor by the time we need them. Any
1271 // that weren't original present will have been instantiated earlier
1273 Instruction* PREInstr = BI->clone();
1274 bool success = true;
1275 for (unsigned i = 0; i < BI->getNumOperands(); ++i) {
1276 Value* op = BI->getOperand(i);
1277 if (isa<Argument>(op) || isa<Constant>(op) || isa<GlobalValue>(op))
1278 PREInstr->setOperand(i, op);
1280 Value* V = lookupNumber(PREPred, VN.lookup(op));
1285 PREInstr->setOperand(i, V);
1289 // Fail out if we encounter an operand that is not available in
1290 // the PRE predecessor. This is typically because of loads which
1291 // are not value numbered precisely.
1298 PREInstr->insertBefore(PREPred->getTerminator());
1299 PREInstr->setName(BI->getName() + ".pre");
1300 predMap[PREPred] = PREInstr;
1301 VN.add(PREInstr, valno);
1304 // Update the availability map to include the new instruction.
1305 localAvail[PREPred]->table.insert(std::make_pair(valno, PREInstr));
1307 // Create a PHI to make the value available in this block.
1308 PHINode* Phi = PHINode::Create(BI->getType(),
1309 BI->getName() + ".pre-phi",
1310 CurrentBlock->begin());
1311 for (pred_iterator PI = pred_begin(CurrentBlock),
1312 PE = pred_end(CurrentBlock); PI != PE; ++PI)
1313 Phi->addIncoming(predMap[*PI], *PI);
1316 localAvail[CurrentBlock]->table[valno] = Phi;
1318 BI->replaceAllUsesWith(Phi);
1321 Instruction* erase = BI;
1323 erase->eraseFromParent();
1329 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
1330 I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
1331 SplitCriticalEdge(I->first, I->second, this);
1333 return changed || toSplit.size();
1336 // iterateOnFunction - Executes one iteration of GVN
1337 bool GVN::iterateOnFunction(Function &F) {
1338 DominatorTree &DT = getAnalysis<DominatorTree>();
1340 cleanupGlobalSets();
1342 // Top-down walk of the dominator tree
1343 bool changed = false;
1344 for (df_iterator<DomTreeNode*> DI = df_begin(DT.getRootNode()),
1345 DE = df_end(DT.getRootNode()); DI != DE; ++DI)
1346 changed |= processBlock(*DI);
1351 void GVN::cleanupGlobalSets() {
1355 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
1356 I = localAvail.begin(), E = localAvail.end(); I != E; ++I)